ISF: Statistical and Deep Learning of High-Resolution Rainfall for Midwest Urban Sustainability Study
DUIRI - Discovery Undergraduate Interdisciplinary Research Internship
Fall 2023
Accepted
climate change and extreme weather, applied and computational mathematics, statistics and data science
As cities grow and more people move into them, they become more vulnerable to the impacts of climate change and extreme weather. This is particularly evident in the Midwest region of the US, including the cities of Chicago and Indianapolis, which have seen an increase in flooding due to more frequent and intense rainfalls. This is a result of more buildings and roads that can't absorb water, leading to rapid runoff that overwhelms the city's drainage systems. One of the ways we can deal with this issue is by building what are called green infrastructures (GIs). These are designed to hold onto rainwater and reduce runoff. But designing these GIs isn't straightforward because it requires detailed information on rainfall patterns which is currently lacking.
In our research, we are trying to fill in this data gap. We're aiming to create detailed maps (O(100m x 100m)) of rainfall in Chicago and Indianapolis, at the neighborhood level, for the years 2005 and 2015. However, the data we need to do this is limited. So, we are going to combine different types of data from various sources, including ground-based measurements, satellite images, and weather model data, to create these maps and answer five important questions:
1. Are the data from different sources consistent with each other?
2. How can we define the 'ground truth' or the most accurate measure of rainfall?
3. How can we best combine different types of data to create a detailed rainfall map?
4. Can adding other types of information, such as measurements of moisture in the atmosphere, improve our maps?
5. Did rainfall patterns, especially the heavy rainfalls that can cause flooding, change significantly over time between 2005 and 2015?
By answering these questions, we hope to improve the way we design GIs and help make Midwest cities more resilient to the impacts of extreme weather as climate changes.
Di Qi
Wen-wen Tung
The project will include a comprehensive interplay between theoretical mathematical modeling, development of computational algorithms and analysis of different types of observational data. First, prototype mathematical models as a system of ODEs and PDEs will be developed to achieve a basic understanding of the fundamental properties of the data structures. Then, data-driven strategies will be used to provide effective formulation for the complex coupled processes from the data. Data-driven models and machine learning tools provide a promising way to address the various difficulties involving the heterogeneous data and complex dynamical models. The proposed data-driven methods will be first tested on the previous simple mathematical models, and next generalized to the realistic data set.
The data in our study include various types of rainfall observations over the greater Chicago or Indianapolis areas. We approach the inconsistency across data from different platforms with statistical calibration. Then, we will downscale the various sources of gridded data via deep learning to the desired high resolution. We will also check if adding measures of atmospheric moisture can improve our model. Finally, we'll compare rainfall data from 2005 to 2010 to identify any changes over time.
The students are expected to have taken basic courses on multivariate calculus (including vector calculus) and linear algebra, statistics, programing. Experiences on ordinary and partial differential equations, machine learning, and atmosphere and ocean science are preferred but not necessary.
3
10 (estimated)